Skip to content
forked from lvdmaaten/bhtsne

Fork of Barnes-Hut t-SNE with improved performance and OpenMP parallelization

License

Notifications You must be signed in to change notification settings

maximsch2/bhtsne

 
 

Repository files navigation

This software package contains a Barnes-Hut implementation of the t-SNE algorithm. The implementation is described in this paper.

Installation

On Linux or OS X, cd to the project dir and compile the source using make. It will generate an executable in <project-dir>/out/bh_tsne.

On Windows using Visual C++, do the following in your command line:

  • Find the vcvars64.bat file in your Visual C++ installation directory. This file may be named vcvars64.bat or something similar. For example:
  // Visual Studio 12
  "C:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\bin\amd64\vcvars64.bat"

  // Visual Studio 2013 Express:
  C:\VisualStudioExp2013\VC\bin\x86_amd64\vcvarsx86_amd64.bat
  • From cmd.exe, go to the directory containing that .bat file and run it.

  • Go to bhtsne directory and run:

  nmake -f Makefile.win all

The executable will be called windows\bh_tsne.exe.

Usage

The code comes with wrappers for Matlab and Python. These wrappers write your data to a file called data.dat, run the bh_tsne binary, and read the result file result.dat that the binary produces. There are also external wrappers available for Torch, R, and Julia. Writing your own wrapper should be straightforward; please refer to one of the existing wrappers for the format of the data and result files.

Demonstration of usage in Matlab:

filename = websave('mnist_train.mat', 'https://github.com/awni/cs224n-pa4/blob/master/Simple_tSNE/mnist_train.mat?raw=true');
load(filename);
numDims = 2; pcaDims = 50; perplexity = 50; theta = .5; alg = 'svd';
map = fast_tsne(digits', numDims, pcaDims, perplexity, theta, alg);
gscatter(map(:,1), map(:,2), labels');

About

Fork of Barnes-Hut t-SNE with improved performance and OpenMP parallelization

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • C++ 78.1%
  • Python 14.7%
  • MATLAB 6.6%
  • Makefile 0.6%